Image to help understand the article
DeepSeek touts faster, more efficient AI model
DeepSeek released DeepSeek-V4.1-Flash-1017 on Oct. 17, 2026, saying the artificial intelligence model improves benchmark performance and completes tasks faster than its predecessor while producing shorter responses.
On DeepSWE 1.1, a benchmark focused on software engineering tasks, the model scored 77%, according to DeepSeek. The company estimated an average cost of about 20 cents per task. The previous DeepSeek-V4.1-Flash scored 74.2% at an estimated 50 cents per run.
DeepSeek also reported scores of 60.5% on AutomationBench and 72.1% on Apex Bench. It said the model outperformed systems including Anthropic’s Claude Opus 5, OpenAI’s GPT-5.6-Sol, Kimi K3 and GLM 5.3 in its coding and agent evaluations. The comparison models were listed as costing from $2 to more than $10 per task.
Those results have not yet been broadly confirmed through independent testing. Benchmarks also may not reflect the quality of code developers encounter in real-world projects. DeepSeek presented its official comparison as an image rather than a data table, making the underlying figures less accessible for analysis.
Wu Yu, DeepSeek’s head of post-training, described the release as the start of an era of “democratized, wish-driven coding.” The model’s key-value cache, a technical component used to speed AI text generation, is listed at 890 bytes per token.
Independent rankings offer a more mixed picture
BenchLM, an independent benchmark aggregator, gave DeepSeek V4.1 Flash 1017 an overall score of 65.61 out of 100 in an Oct. 17 update, ranking it 35th among 954 models. Its results included 90.6% on Terminal-Bench 2.1, 88.1% on CyberGym, 68.9% on AA AutomationBench, 77% on DeepSWE, 68.2% on NL2Repo and 90.9% on GPQA and GPQA-D.
BenchLM also reported a Codeforces score of 3471.0, a FrontierMath v2 Tiers 1-3 result of 67.368%, an Apex score of 65.6% and an AA-LCR score of 84%. The overall index placed the model ahead of DeepSeek V4 Pro 0813, which scored 63.19, and DeepSeek V3.2, which scored 50.2.
The Apex result highlights the uncertainty surrounding benchmark comparisons: BenchLM listed 65.6%, while another comparison cited a 72.1% score. Vals AI did not have a dedicated evaluation page for the 1017 release. Its evaluation of the earlier DeepSeek V4.1 Flash gave that model a Vals Index score of 51.32%, with a margin of error of 1.13 percentage points, ranking it 26th among 49 models.
Pricing targets high-volume developers
DeepSeek-V4.1-Flash-1017 is available through the company’s website, app and application programming interface under the “deepseek-flash” model ID. The service supports a 1 million-token context window, outputs of up to 384,000 tokens and a concurrency limit of 2,500. DeepSeek’s separate “deepseek-v4-pro” API model remains connected to DeepSeek-V4-Pro-0813.
For every 1 million tokens, cached input costs $0.003 during off-peak hours and $0.006 during peak hours. Uncached input costs $0.15 off-peak and 30 cents at peak, while output costs 60 cents off-peak and $1.20 at peak. Peak pricing applies on weekdays from 1 a.m. to 4 a.m. UTC and from 6 a.m. to 10 a.m. UTC.
DeepSeek’s server-side Search and Fetch API tools are free through 4 p.m. UTC on Oct. 31, 2026. After that, each tool will cost 75 cents per 1,000 successful requests.
Competition with U.S. AI companies
DeepSeek’s comparisons place the Chinese developer directly against U.S. companies OpenAI and Anthropic, particularly in the market for AI systems that can write code and carry out multistep tasks. The company’s central pitch is not simply higher benchmark scores but a lower estimated cost per completed job. Still, the conflicting Apex figures and lack of extensive independent testing make it too early to conclude that the model consistently surpasses its American rivals.
Results from the previous DeepSeek-V4.1-Flash showed why broader testing matters. That model narrowly beat Claude Opus 5 and GPT-5.6-Sol on DeepSWE 1.1, but Claude Opus 5 performed substantially better on Terminal-Bench 3.0 and 4.0.
Open weights come with technical caveats
DeepSeek has released the model’s open weights through Hugging Face and ModelScope. It is optimized for DeepSeek Harness version 0.2.1-rc.1 or later, as well as Tencent’s CodeBuddy and WorkBuddy tools.
The earlier V4.1 Flash release used a 552 billion-parameter mixture-of-experts backbone, a 196 billion-parameter conditional memory module and a 1 million-token context window. It activated 8 billion parameters during the initial processing stage and 16 billion during generation, and it included native vision capabilities.
DeepSeek previously warned that sparse-attention errors and state reconstruction could reduce performance in untested situations. It also said the maximum reasoning setting could consume about 2.5 times as many output tokens and that prompts tailored to one production environment might stop working after a model migration. The latest release also arrives amid reports that the Chinese government is considering tighter rules for publishing open-weight models, including possible restrictions on releasing them publicly.
Comments
Post a Comment